1 //===- GlobalMerge.cpp - Internal globals merging -------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This pass merges globals with internal linkage into one. This way all the
10 // globals which were merged into a biggest one can be addressed using offsets
11 // from the same base pointer (no need for separate base pointer for each of the
12 // global). Such a transformation can significantly reduce the register pressure
13 // when many globals are involved.
15 // For example, consider the code which touches several global variables at
18 // static int foo[N], bar[N], baz[N];
20 // for (i = 0; i < N; ++i) {
21 // foo[i] = bar[i] * baz[i];
24 // On ARM the addresses of 3 arrays should be kept in the registers, thus
25 // this code has quite large register pressure (loop body):
32 // Pass converts the code to something like:
40 // for (i = 0; i < N; ++i) {
41 // merged.foo[i] = merged.bar[i] * merged.baz[i];
44 // and in ARM code this becomes:
51 // note that we saved 2 registers here almostly "for free".
53 // However, merging globals can have tradeoffs:
54 // - it confuses debuggers, tools, and users
55 // - it makes linker optimizations less useful (order files, LOHs, ...)
56 // - it forces usage of indexed addressing (which isn't necessarily "free")
57 // - it can increase register pressure when the uses are disparate enough.
59 // We use heuristics to discover the best global grouping we can (cf cl::opts).
61 // ===---------------------------------------------------------------------===//
63 #include "llvm/ADT/BitVector.h"
64 #include "llvm/ADT/DenseMap.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/SmallVector.h"
67 #include "llvm/ADT/Statistic.h"
68 #include "llvm/ADT/StringRef.h"
69 #include "llvm/ADT/Triple.h"
70 #include "llvm/ADT/Twine.h"
71 #include "llvm/CodeGen/Passes.h"
72 #include "llvm/IR/BasicBlock.h"
73 #include "llvm/IR/Constants.h"
74 #include "llvm/IR/DataLayout.h"
75 #include "llvm/IR/DerivedTypes.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/GlobalAlias.h"
78 #include "llvm/IR/GlobalValue.h"
79 #include "llvm/IR/GlobalVariable.h"
80 #include "llvm/IR/Instruction.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/Type.h"
83 #include "llvm/IR/Use.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/InitializePasses.h"
86 #include "llvm/MC/SectionKind.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/Casting.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Target/TargetLoweringObjectFile.h"
93 #include "llvm/Target/TargetMachine.h"
101 using namespace llvm;
103 #define DEBUG_TYPE "global-merge"
105 // FIXME: This is only useful as a last-resort way to disable the pass.
107 EnableGlobalMerge("enable-global-merge", cl::Hidden,
108 cl::desc("Enable the global merge pass"),
111 static cl::opt<unsigned>
112 GlobalMergeMaxOffset("global-merge-max-offset", cl::Hidden,
113 cl::desc("Set maximum offset for global merge pass"),
116 static cl::opt<bool> GlobalMergeGroupByUse(
117 "global-merge-group-by-use", cl::Hidden,
118 cl::desc("Improve global merge pass to look at uses"), cl::init(true));
120 static cl::opt<bool> GlobalMergeIgnoreSingleUse(
121 "global-merge-ignore-single-use", cl::Hidden,
122 cl::desc("Improve global merge pass to ignore globals only used alone"),
126 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
127 cl::desc("Enable global merge pass on constants"),
130 // FIXME: this could be a transitional option, and we probably need to remove
131 // it if only we are sure this optimization could always benefit all targets.
132 static cl::opt<cl::boolOrDefault>
133 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
134 cl::desc("Enable global merge pass on external linkage"));
136 STATISTIC(NumMerged, "Number of globals merged");
140 class GlobalMerge : public FunctionPass {
141 const TargetMachine *TM = nullptr;
143 // FIXME: Infer the maximum possible offset depending on the actual users
144 // (these max offsets are different for the users inside Thumb or ARM
145 // functions), see the code that passes in the offset in the ARM backend
146 // for more information.
149 /// Whether we should try to optimize for size only.
150 /// Currently, this applies a dead simple heuristic: only consider globals
151 /// used in minsize functions for merging.
152 /// FIXME: This could learn about optsize, and be used in the cost model.
153 bool OnlyOptimizeForSize = false;
155 /// Whether we should merge global variables that have external linkage.
156 bool MergeExternalGlobals = false;
160 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
161 Module &M, bool isConst, unsigned AddrSpace) const;
163 /// Merge everything in \p Globals for which the corresponding bit
164 /// in \p GlobalSet is set.
165 bool doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
166 const BitVector &GlobalSet, Module &M, bool isConst,
167 unsigned AddrSpace) const;
169 /// Check if the given variable has been identified as must keep
170 /// \pre setMustKeepGlobalVariables must have been called on the Module that
172 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
173 return MustKeepGlobalVariables.count(GV);
176 /// Collect every variables marked as "used" or used in a landing pad
177 /// instruction for this Module.
178 void setMustKeepGlobalVariables(Module &M);
180 /// Collect every variables marked as "used"
181 void collectUsedGlobalVariables(Module &M, StringRef Name);
183 /// Keep track of the GlobalVariable that must not be merged away
184 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
187 static char ID; // Pass identification, replacement for typeid.
189 explicit GlobalMerge()
190 : FunctionPass(ID), MaxOffset(GlobalMergeMaxOffset) {
191 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
194 explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset,
195 bool OnlyOptimizeForSize, bool MergeExternalGlobals)
196 : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
197 OnlyOptimizeForSize(OnlyOptimizeForSize),
198 MergeExternalGlobals(MergeExternalGlobals) {
199 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
202 bool doInitialization(Module &M) override;
203 bool runOnFunction(Function &F) override;
204 bool doFinalization(Module &M) override;
206 StringRef getPassName() const override { return "Merge internal globals"; }
208 void getAnalysisUsage(AnalysisUsage &AU) const override {
209 AU.setPreservesCFG();
210 FunctionPass::getAnalysisUsage(AU);
214 } // end anonymous namespace
216 char GlobalMerge::ID = 0;
218 INITIALIZE_PASS(GlobalMerge, DEBUG_TYPE, "Merge global variables", false, false)
220 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
221 Module &M, bool isConst, unsigned AddrSpace) const {
222 auto &DL = M.getDataLayout();
223 // FIXME: Find better heuristics
225 Globals, [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
226 return DL.getTypeAllocSize(GV1->getValueType()) <
227 DL.getTypeAllocSize(GV2->getValueType());
230 // If we want to just blindly group all globals together, do so.
231 if (!GlobalMergeGroupByUse) {
232 BitVector AllGlobals(Globals.size());
234 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
237 // If we want to be smarter, look at all uses of each global, to try to
238 // discover all sets of globals used together, and how many times each of
239 // these sets occurred.
241 // Keep this reasonably efficient, by having an append-only list of all sets
242 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
243 // code (currently, a Function) to the set of globals seen so far that are
244 // used together in that unit (GlobalUsesByFunction).
246 // When we look at the Nth global, we know that any new set is either:
247 // - the singleton set {N}, containing this global only, or
248 // - the union of {N} and a previously-discovered set, containing some
249 // combination of the previous N-1 globals.
250 // Using that knowledge, when looking at the Nth global, we can keep:
251 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
252 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
253 // if it actually occurs.
255 // We keep track of the sets of globals used together "close enough".
256 struct UsedGlobalSet {
258 unsigned UsageCount = 1;
260 UsedGlobalSet(size_t Size) : Globals(Size) {}
263 // Each set is unique in UsedGlobalSets.
264 std::vector<UsedGlobalSet> UsedGlobalSets;
266 // Avoid repeating the create-global-set pattern.
267 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
268 UsedGlobalSets.emplace_back(Globals.size());
269 return UsedGlobalSets.back();
272 // The first set is the empty set.
273 CreateGlobalSet().UsageCount = 0;
275 // We define "close enough" to be "in the same function".
276 // FIXME: Grouping uses by function is way too aggressive, so we should have
277 // a better metric for distance between uses.
278 // The obvious alternative would be to group by BasicBlock, but that's in
279 // turn too conservative..
280 // Anything in between wouldn't be trivial to compute, so just stick with
281 // per-function grouping.
283 // The value type is an index into UsedGlobalSets.
284 // The default (0) conveniently points to the empty set.
285 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
287 // Now, look at each merge-eligible global in turn.
289 // Keep track of the sets we already encountered to which we added the
291 // Each element matches the same-index element in UsedGlobalSets.
292 // This lets us efficiently tell whether a set has already been expanded to
293 // include the current global.
294 std::vector<size_t> EncounteredUGS;
296 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
297 GlobalVariable *GV = Globals[GI];
299 // Reset the encountered sets for this global...
300 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
301 // ...and grow it in case we created new sets for the previous global.
302 EncounteredUGS.resize(UsedGlobalSets.size());
304 // We might need to create a set that only consists of the current global.
305 // Keep track of its index into UsedGlobalSets.
306 size_t CurGVOnlySetIdx = 0;
308 // For each global, look at all its Uses.
309 for (auto &U : GV->uses()) {
310 // This Use might be a ConstantExpr. We're interested in Instruction
311 // users, so look through ConstantExpr...
313 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
316 UI = &*CE->use_begin();
318 } else if (isa<Instruction>(U.getUser())) {
325 // ...to iterate on all the instruction users of the global.
326 // Note that we iterate on Uses and not on Users to be able to getNext().
327 for (; UI != UE; UI = UI->getNext()) {
328 Instruction *I = dyn_cast<Instruction>(UI->getUser());
332 Function *ParentFn = I->getParent()->getParent();
334 // If we're only optimizing for size, ignore non-minsize functions.
335 if (OnlyOptimizeForSize && !ParentFn->hasMinSize())
338 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
340 // If this is the first global the basic block uses, map it to the set
341 // consisting of this global only.
343 // If that set doesn't exist yet, create it.
344 if (!CurGVOnlySetIdx) {
345 CurGVOnlySetIdx = UsedGlobalSets.size();
346 CreateGlobalSet().Globals.set(GI);
348 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
351 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
355 // If we already encountered this BB, just increment the counter.
356 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
357 ++UsedGlobalSets[UGSIdx].UsageCount;
361 // If not, the previous set wasn't actually used in this function.
362 --UsedGlobalSets[UGSIdx].UsageCount;
364 // If we already expanded the previous set to include this global, just
365 // reuse that expanded set.
366 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
367 ++UsedGlobalSets[ExpandedIdx].UsageCount;
368 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
372 // If not, create a new set consisting of the union of the previous set
373 // and this global. Mark it as encountered, so we can reuse it later.
374 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
375 UsedGlobalSets.size();
377 UsedGlobalSet &NewUGS = CreateGlobalSet();
378 NewUGS.Globals.set(GI);
379 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
384 // Now we found a bunch of sets of globals used together. We accumulated
385 // the number of times we encountered the sets (i.e., the number of blocks
386 // that use that exact set of globals).
388 // Multiply that by the size of the set to give us a crude profitability
390 llvm::stable_sort(UsedGlobalSets,
391 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
392 return UGS1.Globals.count() * UGS1.UsageCount <
393 UGS2.Globals.count() * UGS2.UsageCount;
396 // We can choose to merge all globals together, but ignore globals never used
397 // with another global. This catches the obviously non-profitable cases of
398 // having a single global, but is aggressive enough for any other case.
399 if (GlobalMergeIgnoreSingleUse) {
400 BitVector AllGlobals(Globals.size());
401 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
402 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
403 if (UGS.UsageCount == 0)
405 if (UGS.Globals.count() > 1)
406 AllGlobals |= UGS.Globals;
408 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
411 // Starting from the sets with the best (=biggest) profitability, find a
413 // The ideal (and expensive) solution can only be found by trying all
414 // combinations, looking for the one with the best profitability.
415 // Don't be smart about it, and just pick the first compatible combination,
416 // starting with the sets with the best profitability.
417 BitVector PickedGlobals(Globals.size());
418 bool Changed = false;
420 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
421 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
422 if (UGS.UsageCount == 0)
424 if (PickedGlobals.anyCommon(UGS.Globals))
426 PickedGlobals |= UGS.Globals;
427 // If the set only contains one global, there's no point in merging.
428 // Ignore the global for inclusion in other sets though, so keep it in
430 if (UGS.Globals.count() < 2)
432 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
438 bool GlobalMerge::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
439 const BitVector &GlobalSet, Module &M, bool isConst,
440 unsigned AddrSpace) const {
441 assert(Globals.size() > 1);
443 Type *Int32Ty = Type::getInt32Ty(M.getContext());
444 Type *Int8Ty = Type::getInt8Ty(M.getContext());
445 auto &DL = M.getDataLayout();
447 LLVM_DEBUG(dbgs() << " Trying to merge set, starts with #"
448 << GlobalSet.find_first() << "\n");
450 bool Changed = false;
451 ssize_t i = GlobalSet.find_first();
454 uint64_t MergedSize = 0;
455 std::vector<Type*> Tys;
456 std::vector<Constant*> Inits;
457 std::vector<unsigned> StructIdxs;
459 bool HasExternal = false;
460 StringRef FirstExternalName;
463 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
464 Type *Ty = Globals[j]->getValueType();
466 // Make sure we use the same alignment AsmPrinter would use.
467 Align Alignment = DL.getPreferredAlign(Globals[j]);
468 unsigned Padding = alignTo(MergedSize, Alignment) - MergedSize;
469 MergedSize += Padding;
470 MergedSize += DL.getTypeAllocSize(Ty);
471 if (MergedSize > MaxOffset) {
475 Tys.push_back(ArrayType::get(Int8Ty, Padding));
476 Inits.push_back(ConstantAggregateZero::get(Tys.back()));
480 Inits.push_back(Globals[j]->getInitializer());
481 StructIdxs.push_back(CurIdx++);
483 MaxAlign = std::max(MaxAlign, Alignment);
485 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
487 FirstExternalName = Globals[j]->getName();
491 // Exit early if there is only one global to merge.
492 if (Tys.size() < 2) {
497 // If merged variables doesn't have external linkage, we needn't to expose
498 // the symbol after merging.
499 GlobalValue::LinkageTypes Linkage = HasExternal
500 ? GlobalValue::ExternalLinkage
501 : GlobalValue::InternalLinkage;
502 // Use a packed struct so we can control alignment.
503 StructType *MergedTy = StructType::get(M.getContext(), Tys, true);
504 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
506 // On Darwin external linkage needs to be preserved, otherwise
507 // dsymutil cannot preserve the debug info for the merged
508 // variables. If they have external linkage, use the symbol name
509 // of the first variable merged as the suffix of global symbol
510 // name. This avoids a link-time naming conflict for the
511 // _MergedGlobals symbols.
513 (IsMachO && HasExternal)
514 ? "_MergedGlobals_" + FirstExternalName
516 auto MergedLinkage = IsMachO ? Linkage : GlobalValue::PrivateLinkage;
517 auto *MergedGV = new GlobalVariable(
518 M, MergedTy, isConst, MergedLinkage, MergedInit, MergedName, nullptr,
519 GlobalVariable::NotThreadLocal, AddrSpace);
521 MergedGV->setAlignment(MaxAlign);
522 MergedGV->setSection(Globals[i]->getSection());
524 const StructLayout *MergedLayout = DL.getStructLayout(MergedTy);
525 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k), ++idx) {
526 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
527 std::string Name(Globals[k]->getName());
528 GlobalValue::VisibilityTypes Visibility = Globals[k]->getVisibility();
529 GlobalValue::DLLStorageClassTypes DLLStorage =
530 Globals[k]->getDLLStorageClass();
532 // Copy metadata while adjusting any debug info metadata by the original
533 // global's offset within the merged global.
534 MergedGV->copyMetadata(Globals[k],
535 MergedLayout->getElementOffset(StructIdxs[idx]));
538 ConstantInt::get(Int32Ty, 0),
539 ConstantInt::get(Int32Ty, StructIdxs[idx]),
542 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
543 Globals[k]->replaceAllUsesWith(GEP);
544 Globals[k]->eraseFromParent();
546 // When the linkage is not internal we must emit an alias for the original
547 // variable name as it may be accessed from another object. On non-Mach-O
548 // we can also emit an alias for internal linkage as it's safe to do so.
549 // It's not safe on Mach-O as the alias (and thus the portion of the
550 // MergedGlobals variable) may be dead stripped at link time.
551 if (Linkage != GlobalValue::InternalLinkage || !IsMachO) {
552 GlobalAlias *GA = GlobalAlias::create(Tys[StructIdxs[idx]], AddrSpace,
553 Linkage, Name, GEP, &M);
554 GA->setVisibility(Visibility);
555 GA->setDLLStorageClass(DLLStorage);
567 void GlobalMerge::collectUsedGlobalVariables(Module &M, StringRef Name) {
568 // Extract global variables from llvm.used array
569 const GlobalVariable *GV = M.getGlobalVariable(Name);
570 if (!GV || !GV->hasInitializer()) return;
572 // Should be an array of 'i8*'.
573 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
575 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
576 if (const GlobalVariable *G =
577 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
578 MustKeepGlobalVariables.insert(G);
581 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
582 collectUsedGlobalVariables(M, "llvm.used");
583 collectUsedGlobalVariables(M, "llvm.compiler.used");
585 for (Function &F : M) {
586 for (BasicBlock &BB : F) {
587 Instruction *Pad = BB.getFirstNonPHI();
591 // Keep globals used by landingpads and catchpads.
592 for (const Use &U : Pad->operands()) {
593 if (const GlobalVariable *GV =
594 dyn_cast<GlobalVariable>(U->stripPointerCasts()))
595 MustKeepGlobalVariables.insert(GV);
601 bool GlobalMerge::doInitialization(Module &M) {
602 if (!EnableGlobalMerge)
605 IsMachO = Triple(M.getTargetTriple()).isOSBinFormatMachO();
607 auto &DL = M.getDataLayout();
608 DenseMap<std::pair<unsigned, StringRef>, SmallVector<GlobalVariable *, 16>>
609 Globals, ConstGlobals, BSSGlobals;
610 bool Changed = false;
611 setMustKeepGlobalVariables(M);
613 // Grab all non-const globals.
614 for (auto &GV : M.globals()) {
615 // Merge is safe for "normal" internal or external globals only
616 if (GV.isDeclaration() || GV.isThreadLocal() || GV.hasImplicitSection())
619 // It's not safe to merge globals that may be preempted
620 if (TM && !TM->shouldAssumeDSOLocal(M, &GV))
623 if (!(MergeExternalGlobals && GV.hasExternalLinkage()) &&
624 !GV.hasInternalLinkage())
627 PointerType *PT = dyn_cast<PointerType>(GV.getType());
628 assert(PT && "Global variable is not a pointer!");
630 unsigned AddressSpace = PT->getAddressSpace();
631 StringRef Section = GV.getSection();
633 // Ignore all 'special' globals.
634 if (GV.getName().startswith("llvm.") ||
635 GV.getName().startswith(".llvm."))
638 // Ignore all "required" globals:
639 if (isMustKeepGlobalVariable(&GV))
642 Type *Ty = GV.getValueType();
643 if (DL.getTypeAllocSize(Ty) < MaxOffset) {
645 TargetLoweringObjectFile::getKindForGlobal(&GV, *TM).isBSS())
646 BSSGlobals[{AddressSpace, Section}].push_back(&GV);
647 else if (GV.isConstant())
648 ConstGlobals[{AddressSpace, Section}].push_back(&GV);
650 Globals[{AddressSpace, Section}].push_back(&GV);
654 for (auto &P : Globals)
655 if (P.second.size() > 1)
656 Changed |= doMerge(P.second, M, false, P.first.first);
658 for (auto &P : BSSGlobals)
659 if (P.second.size() > 1)
660 Changed |= doMerge(P.second, M, false, P.first.first);
662 if (EnableGlobalMergeOnConst)
663 for (auto &P : ConstGlobals)
664 if (P.second.size() > 1)
665 Changed |= doMerge(P.second, M, true, P.first.first);
670 bool GlobalMerge::runOnFunction(Function &F) {
674 bool GlobalMerge::doFinalization(Module &M) {
675 MustKeepGlobalVariables.clear();
679 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
680 bool OnlyOptimizeForSize,
681 bool MergeExternalByDefault) {
682 bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
683 MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
684 return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);